Analogue Neural Network Acceleration

Multiply. Accumulate.
In analogue.

AnalogueAI Labs is developing a patent-pending PWM-based multiply-accumulate IC that performs neural network inference and training entirely in the analogue domain — at a fraction of the energy cost of digital silicon.

PWM·MAC·CELL — CH1/CH2 LIVE
CH1 — Weight PWM CH2 — Data PWM
~10×
Energy reduction vs digital
1T+
MAC ops / watt (projected)
Sky130
Target process node
EP.12
Embodiments filed
01

The PWM Multiply-Accumulate Cell

The fundamental operation in every neural network — the dot product of weights and activations — is a multiply-accumulate (MAC). In digital systems, this is implemented with binary multipliers and adders: fast, precise, but energetically expensive at scale.

Our approach is different. We encode signed numerical values as pulse-width modulated signals, where the temporal position of a pulse within a clock period encodes both magnitude and sign. Multiplication becomes temporal coincidence — the AND of two PWM signals. Accumulation becomes charge integration on a capacitor.

"The sign of the product emerges from the geometry of the timing rather than from explicit digital sign logic."

A rising ramp generates positive values; a falling ramp generates negative values. Both ramps cross VCC/2 at precisely T/2, making opposite-sign signals temporally orthogonal — their AND product is identically zero with no additional logic. An H-bridge accumulator routes same-sign products onto a single integration capacitor with correct polarity, governed by a single XOR gate on two sign latches.

On-chip backpropagation is supported by storing the post-activation output x and recovering the activation derivative σ′(z) = α − x²/α algebraically — without separate storage of pre-activation values.

Signal chain — one MAC cell
Voltage → PWM via bidirectional ramp V→D
Ramp direction encodes sign s_ramp
Pre-bias via current source trim α t_min
AND gate — coincidence detection |w·d|
Leading-edge latch on AND output sign
Inverter chain hold-off 2N·t_inv
XOR of sign latches → H-bridge polarity
H-bridge charges capacitor ∫dt
Tanh activation + σ′ recovery backprop
Σ wᵢ · dᵢ accumulated as V_acc MAC output
02

Research & Intellectual Property

Embodiments 1–9
Unsigned & Mid-Rail Signed PWM MAC
Foundational architecture: voltage-to-PWM conversion, coincidence gate multiplication, capacitive accumulation, and mid-rail encoding for signed values using XOR-based sign correction.
Embodiment 10
Bidirectional Ramp Sign Encoding
Ramp direction encodes sign directly into PWM timing. Temporal orthogonality eliminates XOR gates from the multiply path. Convergence at T/2 provides a universal synchronisation point.
Embodiment 11
H-Bridge Single-Capacitor Accumulator
XOR-routed H-bridge enables signed MAC accumulation on a single integration capacitor. Negative × negative products correctly route forward via the latch polarity signal.
Embodiment 12
Pre-Bias, Leading-Edge Latch & On-Chip Backpropagation
Current source trim α unifies ramp pre-bias, H-bridge settling window, and implicit gradient clipping. Tanh derivative recovered from stored activation output without additional memory.
Circuit Design
Sky130 PDK Implementation
Genetic algorithm optimisation of circuit parameters against SPICE-simulated MNIST accuracy. PMOS current mirror with 4-bit trim DAC. MIM capacitor array sizing for 8-bit precision at 1 MHz PWM.
Training Algorithm
Analogue Backpropagation Protocol
Three-phase MAC accumulation protocol for exact signed MAC. Round-robin storage architecture with sample-and-hold capacitors. Transposed weight access for backward pass without additional memory.
03

Why Analogue PWM?

01
Energy Efficiency
Multiplication by temporal coincidence consumes no switching energy proportional to operand magnitude. A digital multiplier's power scales with bit-width squared; our coincidence gate is a single AND gate whose power is independent of the values being multiplied.
02
Standard CMOS Process
Unlike memristor, PCM, or ferroelectric competitors, our architecture requires no exotic materials or non-standard process steps. It targets the Sky130 open-source PDK — manufacturable at any standard CMOS foundry with no process risk.
03
Inherent Signed Arithmetic
The bidirectional ramp scheme provides fully signed multiply-accumulate without separate sign handling hardware. Temporal orthogonality eliminates mixed-sign products automatically — a mathematical property of the encoding, not a circuit correction.
04
On-Chip Training
The same PWM coincidence gate that performs forward-pass multiplication also computes the activation derivative and backward-pass gradients. Training does not require off-chip computation — the MAC cell is its own backpropagation engine.
04

The Inventors

PG
Peter Goodricke
Co-Inventor · Hardware Architect
Electronics engineer and entrepreneur based in Wellington, New Zealand. Founder of Safety Beacons New Zealand Ltd (emergency beacon electronics design and manufacturing). Deep background in STM32 embedded systems, RF design, SPICE simulation, and KiCad layout. Leads circuit implementation, genetic algorithm optimisation, and Sky130 PDK tapeout preparation.
IW
Ian Wood
Co-Inventor · Mathematical Foundations
Mathematician with expertise in neural network theory and complex-valued signal processing. Provides the formal mathematical framework for the PWM multiply-accumulate architecture, including backpropagation correctness proofs, activation function analysis, and convergence properties of the training protocol.
05

Get in Touch

We are actively seeking strategic investment and research partnerships to advance the Sky130 tapeout and develop the first production array. If you are an investor, foundry partner, or research institution working in edge AI, neuromorphic computing, or low-power inference, we would like to hear from you.

Company AnalogueAI Labs / Safety Beacons New Zealand Ltd
Location Wellington, New Zealand
Web analogue-ai-labs.com
Patents Active provisional application — Embodiments 1–12

Send Enquiry